On learning discriminative embeddings for optimized top-k matching

Optimizing overall classification accuracy in neural networks does not always yield the best top-k accuracy, a critical metric in many real-world applications. This discrepancy is particularly evident in scenarios where multiple classes exhibit high similarity and overlap in the embedding space, leading to class ambiguity during retrieval. Addressing this challenge, the paper proposes a novel method to enhance top-k matching performance by leveraging class relationships in the embedding space. The proposed approach first employs a clustering algorithm to group similar classes into superclusters, capturing their inherent similarity. Next, the compactness of these superclusters is optimized while preserving the discriminative properties of individual classes. This dual optimization improves the separability of classes within superclusters and enhances retrieval accuracy in ambiguous scenarios. Experimental results on diverse datasets, including STL-10, CIFAR-10, CIFAR-100, Stanford Online Products, CARS196, and SCface, demonstrate significant improvements in top-k accuracy, validating the effectiveness and generalizability of the proposed method.